Segmenting thalamic nuclei from manifold projections of multi-contrast MRI

Chang Yan, Muhan Shao, Zhangxing Bian, Anqi Feng, Yuan Xue, Jiachen Zhuo, Rao P. Gullapalli, Aaron Carass, Jerry L. Prince

Abstract

The thalamus is a subcortical gray matter structure that plays a key role in relaying sensory and motor signals within the brain. Its nuclei can atrophy or otherwise be affected by neurological disease and injuries including mild traumatic brain injury. Segmenting both the thalamus and its nuclei is challenging because of the relatively low contrast within and around the thalamus in conventional magnetic resonance (MR) images. This paper explores imaging features to determine key tissue signatures that naturally cluster, from which we can parcellate thalamic nuclei. Tissue contrasts include T1-weighted and T2-weighted images, MR diffusion measurements including FA, mean diffusivity, Knutsson coefficients that represent fiber orientation, and synthetic multi-TI images derived from FGATIR and T1-weighted images. After registration of these contrasts and isolation of the thalamus, we use the uniform manifold approximation and projection (UMAP) method for dimensionality reduction to produce a low-dimensional representation of the data within the thalamus. Manual labeling of the thalamus provides labels for our UMAP embedding from which k nearest neighbors can be used to label new unseen voxels in that same UMAP embedding. N-fold cross-validation of the method reveals comparable performance to state-of-the-art methods for thalamic parcellation.

Introduction

The thalamus acts as the brain's central relay station, processing nearly all sensory and motor information. It is composed of several distinct nuclei, the structural integrity of which is critical; changes in these nuclei are linked to conditions such as Parkinson’s disease, multiple sclerosis, and mild traumatic brain injury (mTBI). However, these sub-structures are notoriously difficult to segment on standard MRI scans due to the low contrast between adjacent nuclei. This study proposes an automated pipeline that integrates diverse imaging data to create a reliable map of these sub-cortical structures.

Methodology

The proposed method treats MRI voxels as high-dimensional feature vectors rather than single intensity values. For each voxel within the thalamus, a feature vector of up to 161 dimensions is constructed, incorporating standard T1 and T2 weights, diffusion metrics (FA, mean diffusivity), fiber orientation coefficients, and synthetic multi-TI images. Spatial coordinates and connectivity maps derived from tractography are also included to provide context.

To manage this high-dimensional space, the researchers utilized Uniform Manifold Approximation and Projection (UMAP). UMAP identifies the underlying manifold of the data and projects it into a low-dimensional (2D, 3D, or 4D) latent space where nuclei naturally cluster.

The UMAP latent space embedding demonstrates how high-dimensional MRI features naturally cluster into distinct thalamic nuclei.

Once embedded in this simplified space, a k-Nearest Neighbor (k-NN) algorithm is employed to assign labels to unseen voxels based on their proximity to manually labeled examples in the latent space.

Experimental Setup

The study utilized data from 44 subjects involved in an mTBI study. Imaging protocols included MPRAGE (T1-w), 3D T2-weighted, DTI, and FGATIR sequences. Ground truth labels were established for 36 subjects by registering the Morel atlas to the scans and manually refining the labels (reducing 19 labels to 13 per hemisphere) to correct for registration errors. A subset of 30 viable subjects was used for a five-fold cross-validation of the pipeline.

Results

The manifold learning approach achieved a volume-weighted Dice score ranging from 0.64 to 0.71, depending on the dimensionality of the latent space. This performance is comparable to current state-of-the-art automated parcellation methods.

Comprehensive visualization of the parcellation pipeline: (a) manual ground truth, (b) 2D UMAP prediction, (c) the 2D latent space embedding, and (d) 4D UMAP prediction.

Key findings regarding the latent space include:

  • Dimensionality Trade-offs: Increasing the latent space from 2D to 4D improved accuracy, particularly for smaller nuclei, but increased training time from approximately 10 minutes to 2 hours.
  • Nuclei-Specific Performance: Larger structures like the Mediodorsal (MD) and Inferior Pulvinar (PuI) were segmented with high accuracy (Dice > 0.80), while smaller, obscure nuclei remained challenging.

Discussion and Conclusion

A significant insight from this work is the importance of spatial context; the most effective feature vectors combined base MRI features with spatial coordinates. Interestingly, the inclusion of cortical connectivity data slightly decreased accuracy. This suggests that the additional dimensions may have introduced noise or that the manual ground truth did not rely on connectivity cues, preventing the model from learning those relationships.

While the results are promising, the study is limited by a relatively small sample size and potential biases in manual labeling. Future work will focus on expanding the dataset to improve robustness and optimizing the computational efficiency of higher-dimensional UMAP embeddings. Ultimately, this manifold learning approach demonstrates that thalamic nuclei are naturally separable in high-dimensional feature space, allowing for accurate parcellation without the need for complex deep learning architectures.

Summary updated

Abstract

The thalamus is a subcortical gray matter structure that plays a key role in relaying sensory and motor signals within the brain. Its nuclei can atrophy or otherwise be affected by neurological disease and injuries including mild traumatic brain injury. Segmenting both the thalamus and its nuclei is challenging because of the relatively low contrast within and around the thalamus in conventional magnetic resonance (MR) images. This paper explores imaging features to determine key tissue signatures that naturally cluster, from which we can parcellate thalamic nuclei. Tissue contrasts include T1-weighted and T2-weighted images, MR diffusion measurements including FA, mean diffusivity, Knutsson coefficients that represent fiber orientation, and synthetic multi-TI images derived from FGATIR and T1-weighted images. After registration of these contrasts and isolation of the thalamus, we use the uniform manifold approximation and projection (UMAP) method for dimensionality reduction to produce a low-dimensional representation of the data within the thalamus. Manual labeling of the thalamus provides labels for our UMAP embedding from which k nearest neighbors can be used to label new unseen voxels in that same UMAP embedding. N -fold cross-validation of the method reveals comparable performance to state-of-the-art methods for thalamic parcellation.
BibTeX
@inproceedings{yan2023segmenting,
  author = {Chang Yan and Muhan Shao and Zhangxing Bian and Anqi Feng and Yuan Xue and Jiachen Zhuo and Rao P. Gullapalli and Aaron Carass and Jerry L. Prince},
  title = {{Segmenting thalamic nuclei from manifold projections of multi-contrast MRI}},
  booktitle = {Image Processing},
  year = {2023},
  doi = {10.1117/12.2654405},
}
Figure